Replication Data for "Randomized Probe Imaging through Deep K-Learning"
-----------------------------------------------------------------------

This dataset contains the key data required to replicate the results in the
paper "Randomized Probe Imaging through Deep K-Learning". This includes:

- the ground truth images used for the experimental and simulated experiments
- the raw simulated and experimental RPI data
- the reconstructed experimental probes and numerical probes used for simulation
- the iterative and deep-k-learning reconstructions

This dataset does not include the trained models themselves, as they take up
significant space. To access the trained models, please contact any of the
authors. This dataset does include the code used to process the data. However,
That code is also available on github at the following link:
https://github.com/zguo0525/Randomized-probe-imaging-through-deep-k-learning/.
The code expects this data to be located in a subfolder, /data/, of the top
level directory.

Finally, because manual moving of data files was a part of the data processing
pipeline we used to generate the results, those wishing to replicate our work
should be aware that some scripts will require particular data files to be
moved into the main folder of the code repository or produce outputs in
different locations from the locations they are stored here. We apologize for
the inconvenience, and please contact the authors if you have any questions.


Folder Organization
===================

There are 3 main folders, one for each of the experiments run in the paper:

- Simulated/R_Sweep
- Simulated/Fixed_R_Noise_Sweep
- Experimental

Each folder contains at least 3 subfolders for the data associated with that
experiment:

- Data (Simulated_Data or RPI_Datasets)
- Iterative_Results
- Neural_Network_Results

Note that the data folders are compressed with 7zip to ensure that the files
stay within the 2GB per-file limit of the Harvard Dataverse.

Within the Iterative_Results folders are the approximants used to train the
networks, as well as the final iterative reconstructions and any intermediate
reconstructions used in figures within the paper

The Neural_Network_Results folders contain the outputs of the various trained
neural networks, using the following naming convention:

- *-not-pretrained-alpha-* files are generative but not pretrained models
- *-pretrained-alpha-* files are generative and pretrained models
- *-not-pretrained-* files are non-generative and not pretrained models
- *-pretrained-* files are non-generative but pretrained models
- *-End-to-end-* files are end-to-end results

In addition, each file is labeled with the R value and mean photons/pixel of
the poisson noise applied (for experimental datasets, this is the "target"
value). Where the photons/pixel is not noted, as in some of the R_sweep data,
the results is from data at 10,000 photons/pixel.

Finally, the ImageNet folder in the top level directory contains the greyscale,
cropped images pulled from the ImageNet dataset that were used for this
expeirment. Where these images were used as the ground truth for the
simulated and experimental portions, these datasets are replicated within the
data folders for the relevant experiment in the appropriate form.